Using protein complexes to predict phenotypic effects of gene mutation

Abstract

Background: Predicting the phenotypic effects of mutations is a central goal of genetics research; it has important applications in elucidating how genotype determines phenotype and in identifying human disease genes.

Results: Using a wide range of functional genomic data from the yeast Saccharomyces cerevisiae, we show that the best predictor of a protein's knockout phenotype is the knockout phenotype of other proteins that are present in a protein complex with it. Even the addition of multiple datasets does not improve upon the predictions made from protein complex membership. Similarly, we find that a proxy for protein complexes is a powerful predictor of disease phenotypes in humans.

Conclusion: We propose that identifying human protein complexes containing known disease genes will be an efficient method for large-scale disease gene discovery, and that yeast may prove to be an informative model system for investigating, and even predicting, the genetic basis of both Mendelian and complex disease phenotypes.

Figure 1

Predictors of phenotype pairs in yeast. (a) Enrichments for phenotype pairs among 20 predictors. An enrichment value of 1 reflects random performance (shown as 'all pairs', the left-most column) and greater than 1 indicates better than random predictive power. Predictors are arranged in order of increasing predictive power. Error bars indicate the hypergeometric standard deviation, which reflects the range of expected variation in the enrichment value. In the inset, red bars are the same as in the main panel a, and are in the same order. Green bars indicate enrichments in the intersection of each dataset with co-expression (r > 0.3). The seven datasets with significant improvements in predictive power are indicated by asterisks. Note that the four co-expression datasets are not counted in the multiple testing correction because they cannot possibly show any improvement when intersecting with another dataset that is a superset. (b) Green bars are the same as in Figure 1a inset. Blue bars indicate the level of enrichment that would be expected by chance, if co-expression was entirely independent of each dataset. Green bars significantly lower than the paired blue bar indicate a dataset that is not independent of co-expression. Error bars indicate the hypergeometric standard deviation, which reflects the range of expected variation in the enrichment value. In the inset, red bars are the same as in panel a and are in the same order as in both panels a and b. Green bars indicate enrichments in the intersection of each dataset with Munich Information Center for Protein Sequences (MIPS) complexes. Note that although many green bars are significantly higher than the paired red bars, no green bars are significantly higher than the MIPS complexes (rightmost) bars. This indicates that no dataset adds to the predictive power of complexes among the set of proteins in MIPS complexes. HTP, high-throughput; LTP, low-throughput; TF, transcription factor.

Figure 2

Predictors of disease gene pairs in human. (a) Enrichments for disease gene pairs among eight predictors. An enrichment value of 1 reflects random performance (shown as 'all pairs', the left-most column). Predictors are arranged in order of increasing predictive power. Error bars indicate the hypergeometric standard deviation, which reflects the range of expected variation in the enrichment value. In the inset red bars are the same as in the main panel a and are in the same order (note the tenfold change in scale). Green bars indicate enrichments in the intersection of each dataset with Human Protein Reference Database (HPRD) interactions. Aside from HPRD intersected with itself or with all pairs, all but one dataset (high-throughput [HTP] interactions) exhibit a significant improvement in predictive power over HPRD interactions alone when intersected with HPRD; this indicates that these datasets are at least partially independent of HPRD. (b) Green bars are the same as in panel a (inset). Blue bars indicate the level of enrichment that would be expected by chance, if HPRD interactions were entirely independent of each dataset. Green bars significantly lower than the paired blue bar indicate a dataset that is not entirely independent of HPRD interactions. The three right-most blue bars are truncated for clarity; their enrichment values are written above each bar. Error bars indicate the hypergeometric standard deviation, which reflects the range of expected variation in the enrichment value.